Testcase: (compiled with -O2 at least)
int f(int a, int b)
{
if (a > 0x7FFFFFF0) return 0;
if (b > 0x7FFFFFF0) return 0;
int c = (a - 20) + (b - 20);
return c > 0x7FFFFFF0;
}
GCC 4.1.2 and 4.3.0 (snapshot from 2006-12-17) optimizes the whole function to
a single "return 0;". This would be correct if the function was actually
written with "c = a + b - 40" under a non-overflow assumption. GCC could indeed
deduce that c is no bigger than 0x7FFFFFFF - 40.
But as the function was originally written, this property does not hold any
longer. For example, a = 0x7FFFFFF0 and b = 41 will not cause any overflow
during computations, and the last conditional shall hence evaluate to true.
The problem is that GCC performs VRP with C language semantic (undefined
behavior on overflow) on code that is no longer the input as written by the
user; so this semantic is not valid at that point. The user input should not
have undergone a transformation based on associativity.
Tested with Debian packages. GCC 3.3.6, 3.4.6, and 4.0.4 generate correct code.
GCC 4.1.2 and 4.3.0 generates wrong code. As the expression "a + b - 40" is
generated early, I suppose any GCC with VRP would produce wrong code for this
testcase.
--
Summary: Wrong variable ranges due to constant folding
Product: gcc
Version: 4.1.2
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: guillaume dot melquiond at ens-lyon dot fr
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30364
